▶ YouTube Summaries

• YT :: https://www.youtube.com/watch?v=jKCjLzjmiaA
• Original title :: Well This Was Unexpected...

OpenAI is terminating its contract with Cursor on 12 November 2026, three months after SpaceX acquired Cursor. Theo walks through the official statement, argues the real trigger is fear that Elon Musk's companies will distill the upcoming Astra model, criticises the public sniping between Cursor and OpenAI leadership, and concludes that the only durable strategy for users is to subscribe directly to the labs rather than routing through a middleman.

What is actually changing

OpenAI notified SpaceX that it will wind down the contract that provides OpenAI models to Cursor, with a proposed shut-off date of 12 November 2026 — the maximum notice the contract allows. Cursor subscriptions will no longer include OpenAI models after that date, and Cursor will never receive Astra, OpenAI's upcoming model.

Workarounds OpenAI explicitly preserves: bring-your-own OpenAI API key inside Cursor, and the Codex IDE extension running inside Cursor. Theo was separately told by an OpenAI contact that third-party harnesses like his own T3 Code are not affected and will keep working.

The SpaceX acquisition is the background

SpaceX acquired Cursor outright rather than taking the originally floated deal — either a $10B partnership for access to Cursor's data and RL engineers, or a $60B acquisition at year end. Theo notes he was an early Cursor investor, still doesn't know whether he'll be able to sell the resulting stock, and considers his allegiance to Cursor over either way.

His read on why Grok improved so sharply: SpaceX bought the data. Cursor's accumulated coding traces are what made models like Composer good, and that pipeline now belongs to Musk.

Why OpenAI pulled the plug: Astra and distillation

The clause that matters is the change-of-control cancellation window. OpenAI says it cannot be confident SpaceX will honour its terms of service, citing Musk's companies breaking contracts before and Musk admitting under oath that xAI distilled OpenAI data.

The decisive detail, which Theo says most coverage missed, is Astra. OpenAI is worried enough about Astra being distilled that it is cutting off any partner not clearly compliant. The economics make it worse: a customer paying Cursor to use OpenAI models means Cursor collects the input/output pairs for free instead of paying API rates to generate them — double dipping, unless the user has opted out of data collection.

Precedent, but not from OpenAI

• Anthropic banned Claude in Windsurf when OpenAI was rumoured to be acquiring it.
• Anthropic banned xAI from reaching Claude through Cursor, early in 2026, over the same distillation worry.
• Anthropic revoked OpenAI's own Claude API access, which is why OpenAI's self-run benchmarks rarely include Anthropic models.

So the move has precedent — just not from OpenAI, which makes it a marker of how bad the OpenAI/Musk relationship has become. Theo's own position is that he criticised Anthropic for this and will criticise OpenAI equally: paying for a model and keeping the outputs is legitimate, and the data SpaceX cannot obtain is only locked away because OpenAI and Anthropic scraped the open web first.

Anthropic, meanwhile, publicly sided with Cursor — its head of compute promised more capacity for Claude in Cursor. Theo reads this as Anthropic's dependence on SpaceX compute plus shared dislike of OpenAI, and notes the awkward timing: a day later Claude Code limits were cut by 17%.

The public spat, and the token-share argument

Thibault at OpenAI posted a reasonable thread listing the workarounds. Cursor's response is where Theo objects, on two counts:

• "block Cursor users from accessing OpenAI models" reads as if Cursor users are banned from OpenAI, which is false — API keys and the Codex extension still work.
• "OpenAI models serve about 5% of Cursor user traffic" is unqualified. Traffic measured in tokens systematically understates OpenAI models because they are far more token-efficient. On Cursor's own bench, Fable scores 70.5% at 103k tokens per task while Soul scores 67.2% at 28k — a 3x gap; artificial analysis shows 36k vs 17k. If the denominator is tokens, 5% is closer to 15% in requests or value.

Thibault then replied demanding Michael show the math. Theo thinks that reply shouldn't have been made: neither executive is personally affected, users are, and public beef between company leadership over measurement methodology just poisons the situation.

What to do about it

There is no good one-subscription-for-every-model option, and Theo says there structurally can't be. Labs run 70–90% inference margins and subsidise their own subscriptions far beyond what a reseller can match — a prompt that costs a third party $2 may cost the lab 20 cents. Cursor floated it partly on the value of the data it kept.

Practical advice:

• If you need frontier performance, pay the labs directly — a $200 Claude Code or Codex subscription dwarfs any aggregator in value.
• Grok 4.5/4.6 are genuinely usable but not frontier; after Fable and Soul they feel like a regression to manual orchestration.
• For a single unified UI, the open-source harnesses are the safer bet: T3 Code, Gene, Superset, Herder.
• Own your tools and keep your relationships with providers direct, not routed through a layer that can be cut off.

Theo expects more moves like this from Anthropic, OpenAI, Google and SpaceX — this is the start, not a one-off. He closes with a tangent at Google: its models are the furthest thing from the company's problem, Antigravity is a bad app, and Google is the most aggressive about banning accounts for using subscriptions elsewhere, which blocks integration into third-party harnesses.